Ai and ML based CSI compression with test decoders

Abstract

A test equipment (TE) may be configured to train a two-sided machine learning model for encoding and decoding channel state information (CSI) that is associated with a cellular network. TE may be configured to obtain training data associated with CSI and the cellular network, input the training data to one or more two-sided machine learning (ML) models, each comprising an encoder and a decoder, compress, by the encoder the channel state information (CSI) resulting in compressed CSI data, and decode, by the decoder the compressed CSI data, to obtain second CSI data.

Description

Al AND ML BASED CSI COMPRESSION WITH TEST DECODERSCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U.S. Provisional Application No. 63 / 703,743, filed October 4, 2024, the content of which is incorporated herein in its entirety for all purposes.TECHNICAL FIELD

[0002] This disclosure relates generally to wireless technology and more particularly to AI / ML testability and interoperability framework for CSI compression with test decoders.BACKGROUND

[0003] Fifth generation mobile network (5G) is a wireless standard that aims to improve upon data transmission speed, reliability, availability, and more. 5G may also be referred to as new radio (NR). The wireless standard includes numerous procedures that may be implemented by a transmitting device or a receiving device that improves the latency, the speed, and the reliability of uplink and downlink transmissions.SUMMARY

[0004] Aspects of the present disclosure relate to 5G new radio (NR) operating in the licensed spectrum or in the shared and unlicensed spectrum (NR-U).

[0005] Aspects in the present disclosure relate to the interoperability and testability aspects for AI / ML for NR air interface for CSI compression and prediction. Aspects of the present disclosure relate to the general framework of AI / ML concerning air interfaces. Aspects described enable the implementation of recommended use cases. Furthermore, several study objectives aim to address outstanding issues, with the goal of enhancing understanding in preparation for additional normative effort.

[0006] In an aspect, a system comprises one or more computing devices, configured to perform operations comprising: obtaining training data associated with channel state information (CSI) and a cellular network; inputting the training data to one or more two-sided machine learning (ML) models, each comprising an encoder and a decoder; compressing, by the encoder the channel state information (CSI) resulting in compressed CSI data; and decoding, by the decoder the compressed CSI data, to obtain second CSI data. The resulting two-sided ML model comprises a trained encoder and trained decoder. This trained decoder14904-9517-2718, v. 3may be integrated into test equipment (TE) that can be configured to test a device under test (DUT) such as a UE.

[0007] In an embodiment, the operations further comprise enforcing at least one of: weight constraints, dropout layers, early stopping criteria, or noise injection. In an embodiment, the training data comprises a data associated with a clustered delay line (CDL) channel or an Urban Macrocell (Uma) channels. This data may be disproportionately large compared to other training data, to prioritize or emphasize CDL or Uma channel behavior for the test encoder.

[0008] In an embodiment, obtaining the training data comprises obtaining portions of the training data from a plurality of data sources, and generating synthetic data from the portions of the training data, wherein the training data comprises a combination of the portions of the training data and the synthetic data.

[0009] In an embodiment, generating the synthetic data comprises applying a generative model or a statistical model to the portions of the training data to generate the synthetic data.

[0010] In an embodiment, the one or more two-sided ML models comprise a plurality of two-sided ML models, each trained with a respective one of a plurality of subsets of the training data. In an embodiment, each of the plurality of the subsets of training data are associated with distinct scenarios comprising different channel conditions, different network conditions, or different combinations of a channel condition and network condition. In an embodiment, the different network conditions comprises at least one of: an indoor propagation environment, an outdoor propagation environment, a location, or a distance. In an embodiment, the different channel conditions comprises at least one of: a different doppler speed, a different carrier frequency, a different channel.

[0011] In an embodiment, a plurality of subsets of the training data are each associated with a training set identifier (ID) that is associated with a respective one of a plurality of ML models trained with that subset of the training data. In such a manner, each ML model and test encoder thereof may be identified based on the training data or subset of training data used to train that test encoder. A DUT may be tested under different network conditions and scenarios with a corresponding test encoder.

[0012] In an embodiment, one of the plurality of ML models is trained by one of the subsets of the training data that is associated with a mix of the distinct scenarios. This may provide a more robust test encoder that is less overfitted to a particular scenario.24904-9517-2718, v. 3

[0013] In an embodiment, the training data comprises multiple datasets, each comprising a range of: different doppler speeds, different carrier frequencies, or different propagation environments (indoor, outdoor).

[0014] In an embodiment, the decoder satisfies a specification defining a model type, a model structure, an activation function, a maximum floating-point operations per seconds (FLOPs), or a maximum number or size of parameters.

[0015] In an embodiment, the encoder of the ML model may used to train a UE encoder.

[0016] In an embodiment, the decoder of the ML model may be deployed to a gNB or used in TE to test a UE.

[0017] In an aspect, a base station comprises a one or more antenna arrays and a processor that is configured to cause the base station to receive compressed CSI and process it with a decoder or copy thereof, of the two-sided ML model.

[0018] In an aspect, a user equipment (UE) comprises a one or more radios and a processor that is configured to cause the UE to apply an encoder to CSI, and transmit the resulting compressed CSI to a gNB, where the encoder is trained as part of the two-sided model described, or a copy thereof.

[0019] In an aspect, a processor (e.g., a baseband processor) of a UE or a base station is configured to perform the operations described. In an aspect, a non-transitory computer- readable memory may store instructions that, when executed by such a processor, causes the UE or the base station to perform the operations described.

[0020] Other technical features may be readily apparent to one skilled in the art from the following figures, descriptions, and claims.BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Aspects of the present disclosure are illustrated by way of example and not limitation in the figures of the accompanying drawings in which like references indicate similar elements.

[0022] FIG. 1 illustrates an example wireless communication system, in accordance with an embodiment.

[0023] FIG. 2 illustrates uplink and downlink communications, in accordance with an embodiment.

[0024] FIG. 3 illustrates an example block diagram of a user equipment (UE), in accordance with an embodiment.34904-9517-2718, v. 3

[0025] FIG. 4 illustrates an example block diagram of a base station (BS), in accordance with an embodiment.

[0026] FIG. 5 illustrates an example block diagram of cellular communication circuitry, in accordance with an embodiment.

[0027] FIG. 6 shows an example of training a test decoder with type 3 training, in accordance with an embodiment.

[0028] FIG. 7 shows an example system for implementing a test decoder, in accordance with an embodiment.

[0029] FIG. 8 illustrations a diagram for identification, training, or testing of Al-based models, in accordance with an embodiment.

[0030] FIG. 9 illustrates an example method for implementing test equipment (TE), in accordance with an embodiment.DETAILED DESCRIPTION

[0031] A method and apparatus is described that relates to a wireless communication and artificial intelligence (AI) / machine learning (ML) testability and interoperability framework for Channel State Information (CSI) compression with partially specified test decoders.

[0032] It will be apparent, however, to one skilled in the art, that aspects of the present disclosure may be practiced without these specific details. In other instances, well-known components (e.g., network and UE components), structures, and techniques have not been shown in detail in order not to obscure the understanding of this description.

[0033] Reference in the specification to “some aspects” or “an aspect” means that a particular feature, structure, or characteristic described in connection with the aspect can be included in at least one aspect of the disclosure. The appearances of the phrase “in some aspects” in various places in the specification do not necessarily all refer to the same aspect.

[0034] In the following description and claims, the terms “coupled” and “connected,” along with their derivatives, may be used. “Coupled” is used to indicate that two or more elements, which may or may not be in direct physical or electrical contact with each other, co-operate or interact with each other. “Connected” is used to indicate the establishment of communication between two or more elements that are coupled with each other.

[0035] The processes depicted in the figures that follow, are performed by processing logic that comprises hardware (e.g., circuitry, dedicated logic, etc.), software (such as is run on a general-purpose computer system or a dedicated machine), or a combination of both.44904-9517-2718, v. 3Although the processes are described below in terms of some sequential operations, it should be appreciated that some of the operations described may be performed in different order. Moreover, some operations may be performed in parallel rather than sequentially.

[0036] The terms “server,” “client,” and “device” are intended to refer generally to data processing systems rather than specifically to a particular form factor for the server, client, and / or device.

[0037] FIG. 1 illustrates a simplified example wireless communication system, according to some aspects. It is noted that the system of FIG. 1 is merely one example of a possible system, and that features of this disclosure may be implemented in any of various systems, as desired.

[0038] As shown, the example wireless communication system includes a base station 102 A which communicates over a transmission medium with one or more user devices 106A, 106B, etc., through 106N. Each of the user devices may be referred to as a “user equipment” (UE).

[0039] The base station (BS) 102A may be a base transceiver station (BTS) or cell site (a “cellular base station”) and may include hardware that enables wireless communication with the UEs 106 A through 106N.

[0040] The communication area (or coverage area) of the base station may be referred to as a “cell.” The base station 102A and the UEs 106 may be configured to communicate over the transmission medium using any of various radio access technologies (RATs), also referred to as wireless communication technologies, or telecommunication standards, such as GSM, UMTS (associated with, for example, WCDMA or TD-SCDMA air interfaces), LTE, LTE- Advanced (LTE- A), 5G new radio (5GNR), HSPA, 3GPP2 CDMA2000 (e g., IxRTT, IxEV- DO, HRPD, eHRPD), etc. Note that if the base station 102A is implemented in the context of LTE, it may alternately be referred to as an ‘eNodeB’ or ‘eNB’. Note that if the base station 102A is implemented in the context of 5G NR, it may alternately be referred to as ‘gNodeB’ or ‘gNB’.

[0041] As shown, the base station 102A may also be equipped to communicate with a network 100 (e.g., a core network of a cellular service provider, a telecommunication network such as a public switched telephone network (PSTN), and / or the Internet, among various possibilities). Thus, the base station 102A may facilitate communication between the user devices and / or between the user devices and the network 100. In particular, the cellular base station 102 A may provide UEs 106 with various telecommunication capabilities, such as voice, SMS and / or data services.54904-9517-2718, v. 3

[0042] Base station 102 A and other similar base stations (such as base stations 102B . . . 102N) operating according to the same or a different cellular communication standard may thus be provided as a network of cells, which may provide continuous or nearly continuous overlapping service to UEs 106A-N and similar devices over a geographic area via one or more cellular communication standards.

[0043] Thus, while base station 102A may act as a “serving cell” for UEs 106A-N as illustrated in FIG. 1, each UE 106 may also be capable of receiving signals from (and possibly within communication range of) one or more other cells (which might be provided by base stations 102B-N and / or any other base stations), which may be referred to as “neighboring cells”. Such cells may also be capable of facilitating communication between user devices and / or between user devices and the network 100. Such cells may include “macro” cells, “micro” cells, “pico” cells, and / or cells which provide any of various other granularities of service area size. For example, base stations 102A-B illustrated in FIG. 1 might be macro cells, while base station 102N might be a micro cell. Other configurations are also possible.

[0044] In some aspects, base station 102A may be a next generation base station, e.g., a 5G New Radio (5G NR) base station, or “gNB”. In some aspects, a gNB may be connected to a legacy evolved packet core (EPC) network and / or to a NR core (NRC) network. In addition, a gNB cell may include one or more transition and reception points (TRPs). In addition, a UE capable of operating according to 5G NR may be connected to one or more TRPs within one or more gNBs.

[0045] Note that a UE 106 may be capable of communicating using multiple wireless communication standards. For example, the UE 106 may be configured to communicate using a wireless networking (e.g., Wi-Fi) and / or peer-to-peer wireless communication protocol (e.g., Bluetooth, Wi-Fi peer-to-peer, etc.) in addition to at least one cellular communication protocol (e.g., GSM, UMTS (associated with, for example, WCDMA or TD-SCDMA air interfaces), LTE, LTE-A, 5G NR, HSPA, 3GPP2 CDMA2000 (e.g, IxRTT, IxEV-DO, HRPD, eHRPD), etc.). The UE 106 may also or alternatively be configured to communicate using one or more global navigational satellite systems (GNSS, e.g., GPS or GLONASS), one or more mobile television broadcasting standards (e.g., ATSC-M / H or DVB-H), and / or any other wireless communication protocol, if desired. Other combinations of wireless communication standards (including more than two wireless communication standards) are also possible.64904-9517-2718, v. 3

[0046] FIG. 2 illustrates UE 106 A that can be in communication with a base station 102 through uplink and downlink communications, according to some aspects. The UEs may each be a device with cellular communication capability such as a mobile phone, a hand-held device, a computer or a tablet, or virtually any type of wireless device.

[0047] The UE may include a processor that is configured to execute program instructions stored in memory. The UE may perform any of the method aspects described herein by executing such stored instructions. Alternatively, or in addition, the UE may include a programmable hardware element such as an FPGA (field-programmable gate array) that is configured to perform any of the method aspects described herein, or any portion of any of the method aspects described herein.

[0048] The UE may include one or more antennas for communicating using one or more wireless communication protocols or technologies. In some aspects, the UE may be configured to communicate using, for example, CDMA2000 (IxRTT / lxEV- DO / HRPD / eHRPD) or LTE using a single shared radio and / or GSM or LTE using the single shared radio. The shared radio may couple to a single antenna, or may couple to multiple antennas (e.g., for MIMO) for performing wireless communications. In general, a radio may include any combination of a baseband processor, analog RF signal processing circuitry (e.g., including filters, mixers, oscillators, amplifiers, etc.), or digital processing circuitry (e.g., for digital modulation as well as other digital processing). Similarly, the radio may implement one or more receive and transmit chains using the aforementioned hardware. For example, the UE 106 may share one or more parts of a receive and / or transmit chain between multiple wireless communication technologies, such as those discussed above.

[0049] In some aspects, the UE may include separate transmit and / or receive chains (e.g., including separate antennas and other radio components) for each wireless communication protocol with which it is configured to communicate. As a further possibility, the UE may include one or more radios which are shared between multiple wireless communication protocols, and one or more radios which are used exclusively by a single wireless communication protocol. For example, the UE might include a shared radio for communicating using either of LTE or 5G NR (or LTE or IxRTTor LTE or GSM), and separate radios for communicating using each of Wi-Fi and Bluetooth. Other configurations are also possible.

[0050] FIG. 3 illustrates an example simplified block diagram of a communication device 106, according to some aspects. It is noted that the block diagram of the communication74904-9517-2718, v. 3device of FIG. 3 is only one example of a possible communication device. According to aspects, communication device 106 may be a UE device, a mobile device or mobile station, a wireless device or wireless station, a desktop computer or computing device, a mobile computing device (e.g., a laptop, notebook, or portable computing device), a tablet and / or a combination of devices, among other devices. As shown, the communication device 106 may include a set of components 300 configured to perform core functions. For example, this set of components may be implemented as a system on chip (SOC), which may include portions for various purposes. Alternatively, this set of components 300 may be implemented as separate components or groups of components for the various purposes. The set of components 300 may be coupled (e.g., communicatively; directly or indirectly) to various other circuits of the communication device 106.

[0051] For example, the communication device 106 may include various types of memory (e.g., including NAND flash 310), an input / output interface such as connector I / F 320 (e.g., for connecting to a computer system; dock; charging station; input devices, such as a microphone, camera, keyboard; output devices, such as speakers; etc.), the display 360, which may be integrated with or external to the communication device 106, and cellular communication circuitry 330 such as for 5GNR, LTE, GSM, etc., and short to medium range wireless communication circuitry 329 (e.g., Bluetooth™ and WLAN circuitry). In some aspects, communication device 106 may include wired communication circuitry (not shown), such as a network interface card, e.g., for Ethernet.

[0052] The cellular communication circuitry 330 may couple (e.g., communicatively; directly or indirectly) to one or more antennas, such as antennas 335 and 336 as shown. The short to medium range wireless communication circuitry 329 may also couple (e.g., communicatively; directly or indirectly) to one or more antennas, such as antennas 337 and 338 as shown. Alternatively, the short to medium range wireless communication circuitry 329 may couple (e.g., communicatively; directly or indirectly) to the antennas 335 and 336 in addition to, or instead of, coupling (e.g., communicatively; directly or indirectly) to the antennas 337 and 338. The short to medium range wireless communication circuitry 329 and / or cellular communication circuitry 330 may include multiple receive chains and / or multiple transmit chains for receiving and / or transmitting multiple spatial streams, such as in a multiple-input multiple output (MIMO) configuration.

[0053] In some aspects, as further described below, cellular communication circuitry 330 may include dedicated receive chains (including and / or coupled to, e.g., communicatively;84904-9517-2718, v. 3directly or indirectly, dedicated processors and / or radios) for multiple radio access technologies (RATs) (e.g., a first receive chain for LTE and a second receive chain for 5G NR). In addition, in some aspects, cellular communication circuitry 330 may include a single transmit chain that may be switched between radios dedicated to specific RATs. For example, a first radio may be dedicated to a first RAT, e.g., LTE, and may be in communication with a dedicated receive chain and a transmit chain shared with an additional radio, e.g., a second radio that may be dedicated to a second RAT, e.g., 5G NR, and may be in communication with a dedicated receive chain and the shared transmit chain.

[0054] The communication device 106 may also include and / or be configured for use with one or more user interface elements. The user interface elements may include any of various elements, such as display 360 (which may be a touchscreen display), a keyboard (which may be a discrete keyboard or may be implemented as part of a touchscreen display), a mouse, a microphone and / or speakers, one or more cameras, one or more buttons, and / or any of various other elements capable of providing information to a user and / or receiving or interpreting user input.

[0055] The communication device 106 may further include one or more smart cards 345 that include SIM (Subscriber Identity Module) functionality, such as one or more UICC(s) (Universal Integrated Circuit Card(s)) cards 345.

[0056] As shown, the SOC 300 may include processor(s) 302, which may execute program instructions for the communication device 106 and display circuitry 304, which may perform graphics processing and provide display signals to the display 360. The processor(s) 302 may also be coupled to memory management unit (MMU) 340, which may be configured to receive addresses from the processor(s) 302 and translate those addresses to locations in memory (e.g., memory 306, read only memory (ROM) 350, NAND flash memory 310) and / or to other circuits or devices, such as the display circuitry 304, short range wireless communication circuitry 229, cellular communication circuitry 330, connector I / F 320, and / or display 360. The MMU 340 may be configured to perform memory protection and page table translation or set up. In some aspects, the MMU 340 may be included as a portion of the processor(s) 302.

[0057] As noted above, the communication device 106 may be configured to communicate using wireless and / or wired communication circuitry. The communication device 106 may also be configured to determine a physical downlink shared channel scheduling resource for a user equipment device and a base station. Further, the communication device 106 may be94904-9517-2718, v. 3configured to group and select CCs from the wireless link and determine a virtual CC from the group of selected CCs. The wireless device may also be configured to perform a physical downlink resource mapping based on an aggregate resource matching patterns of groups of CCs.

[0058] As described herein, the communication device 106 may include hardware and software components for implementing the above features for determining a physical downlink shared channel scheduling resource for a communications device 106 and a base station. The processor 302 of the communication device 106 may be configured to implement part or all of the features described herein, e.g., by executing program instructions stored on a memory medium (e.g., a non-transitory computer-readable memory medium). Alternatively (or in addition), processor 302 may be configured as a programmable hardware element, such as an FPGA (Field Programmable Gate Array), or as an ASIC (Application Specific Integrated Circuit). Alternatively (or in addition) the processor 302 of the communication device 106, in conjunction with one or more of the other components 300, 304, 306, 310, 320, 329, 330, 340, 345, 350, 360 may be configured to implement part or all of the features described herein.

[0059] In addition, as described herein, processor 302 may include one or more processing elements. Thus, processor 302 may include one or more integrated circuits (ICs) that are configured to perform the functions of processor 302. In addition, each integrated circuit may include circuitry (e.g., first circuitry, second circuitry, etc.) configured to perform the functions of processor(s) 302.

[0060] Further, as described herein, cellular communication circuitry 330 and short-range wireless communication circuitry 329 may each include one or more processing elements. In other words, one or more processing elements may be included in cellular communication circuitry 330 and, similarly, one or more processing elements may be included in short range wireless communication circuitry 329. Thus, cellular communication circuitry 330 may include one or more integrated circuits (ICs) that are configured to perform the functions of cellular communication circuitry 330. In addition, each integrated circuit may include circuitry (e.g., first circuitry, second circuitry, etc.) configured to perform the functions of cellular communication circuitry 230. Similarly, the short-range wireless communication circuitry 329 may include one or more ICs that are configured to perform the functions of short-range wireless communication circuitry 32. In addition, each integrated circuit may include circuitry (e.g., first circuitry, second circuitry, etc.) configured to perform the functions of short-range wireless communication circuitry 329.104904-9517-2718, v. 3

[0061] FIG. 4 illustrates an example block diagram of a base station 102, according to some aspects. It is noted that the base station of FIG. 4 is merely one example of a possible base station. As shown, the base station 102 may include processor(s) 404 which may execute program instructions for the base station 102. The processor(s) 404 may also be coupled to memory management unit (MMU) 440, which may be configured to receive addresses from the processor(s) 404 and translate those addresses to locations in memory (e.g., memory 460 and read only memory (ROM) 450) or to other circuits or devices.

[0062] The base station 102 may include at least one network port 470. The network port 470 may be configured to couple to a telephone network and provide a plurality of devices, such as UE devices 106, access to the telephone network as described above in FIGS. 1 and 2.

[0063] The network port 470 (or an additional network port) may also or alternatively be configured to couple to a cellular network, e.g., a core network of a cellular service provider. The core network may provide mobility related services and / or other services to a plurality of devices, such as UE devices 106. In some cases, the network port 470 may couple to a telephone network via the core network, and / or the core network may provide a telephone network (e.g., among other UE devices serviced by the cellular service provider).

[0064] In some aspects, base station 102 may be a next generation base station, e.g., a 5G New Radio (5G NR) base station, or “gNB”. In such aspects, base station 102 may be connected to a legacy evolved packet core (EPC) network and / or to a NR core (NRC) network. In addition, base station 102 may be considered a 5G NR cell and may include one or more transition and reception points (TRPs). In addition, a UE capable of operating according to 5G NR may be connected to one or more TRPs within one or more gNBs. In some aspects, the base station can operate in 5G NR-U mode.

[0065] The base station 102 may include at least one antenna 434, and possibly multiple antennas. The at least one antenna 434 may be configured to operate as a wireless transceiver and may be further configured to communicate with UE devices 106 via radio 430. The antenna 434 communicates with the radio 430 via communication chain 432. Communication chain 432 may be a receive chain, a transmit chain or both. The radio 430 may be configured to communicate via various wireless communication standards, including, but not limited to, 5G NR, 5G NR-U, LTE, LTE-A, GSM, UMTS, CDMA2000, Wi-Fi, etc.

[0066] The base station 102 may be configured to communicate wirelessly using multiple wireless communication standards. In some instances, the base station 102 may include114904-9517-2718, v. 3multiple radios, which may enable the base station 102 to communicate according to multiple wireless communication technologies. For example, as one possibility, the base station 102 may include an LTE radio for performing communication according to LTE as well as a 5G NR radio for performing communication according to 5G NR and 5G NR-U. In such a case, the base station 102 may be capable of operating as both an LTE base station and a 5G NR base station. As another possibility, the base station 102 may include a multi-mode radio which is capable of performing communications according to any of multiple wireless communication technologies (e.g., 5G NR and Wi-Fi, LTE and Wi-Fi, LTE and UMTS, LTE and CDMA2000, UMTS and GSM, etc.).

[0067] As described further subsequently herein, the BS 102 may include hardware and software components for implementing or supporting implementation of features described herein. The processor 404 of the base station 102 may be configured to implement or support implementation of part or all of the methods described herein, e.g., by executing program instructions stored on a memory medium (e.g., a non-transitory computer-readable memory medium). Alternatively, the processor 404 may be configured as a programmable hardware element, such as an FPGA (Field Programmable Gate Array), or as an ASIC (Application Specific Integrated Circuit), or a combination thereof. Alternatively (or in addition) the processor 404 of the BS 102, in conjunction with one or more of the other components 430, 432, 434, 440, 450, 460, 470 may be configured to implement or support implementation of part or all of the features described herein.

[0068] In addition, as described herein, processor(s) 404 may be comprised of one or more processing elements. In other words, one or more processing elements may be included in processor(s) 404. Thus, processor(s) 404 may include one or more integrated circuits (ICs) that are configured to perform the functions of processor(s) 404. In addition, each integrated circuit may include circuitry (e.g., first circuitry, second circuitry, etc.) configured to perform the functions of processor(s) 404.

[0069] Further, as described herein, radio 430 may be comprised of one or more processing elements. In other words, one or more processing elements may be included in radio 430. Thus, radio 430 may include one or more integrated circuits (ICs) that are configured to perform the functions of radio 430. In addition, each integrated circuit may include circuitry (e.g., first circuitry, second circuitry, etc.) configured to perform the functions of radio 430.

[0070] FIG. 5 illustrates an example simplified block diagram of cellular communication circuitry, according to some aspects. It is noted that the block diagram of the cellular124904-9517-2718, v. 3communication circuitry of FIG. 5 is only one example of a possible cellular communication circuit. According to aspects, cellular communication circuitry 330 may be included in a communication device, such as communication device 106 described above. As noted above, communication device 106 may be a user equipment (UE) device, a mobile device or mobile station, a wireless device or wireless station, a desktop computer or computing device, a mobile computing device (e.g., a laptop, notebook, or portable computing device), a tablet and / or a combination of devices, among other devices.

[0071] The cellular communication circuitry 330 may couple (e.g., communicatively; directly or indirectly) to one or more antennas, such as antennas 335 a-b and 336 as shown (in FIG. 3). In some aspects, cellular communication circuitry 330 may include dedicated receive chains (including and / or coupled to, e.g., communicatively; directly or indirectly, dedicated processors and / or radios) for multiple RATs (e.g., a first receive chain for LTE and a second receive chain for 5GNR). For example, as shown in FIG. 5, cellular communication circuitry 330 may include a modem 510 and a modem 520. Modem 510 may be configured for communications according to a first RAT, e.g., such as LTE or LTE-A, and modem 520 may be configured for communications according to a second RAT, e.g., such as 5G NR.

[0072] As shown, modem 510 may include one or more processors 512 and a memory 516 in communication with processors 512. Modem 510 may be in communication with a radio frequency (RF) front end 530. RF front end 530 may include circuitry for transmitting and receiving radio signals. For example, RF front end 530 may include receive circuitry (RX) 532 and transmit circuitry (TX) 534. In some aspects, receive circuitry 532 may be in communication with downlink (DL) front end 550, which may include circuitry for receiving radio signals via antenna 335a.

[0073] Similarly, modem 520 may include one or more processors 522 and a memory 526 in communication with processors 522. Modem 520 may be in communication with an RF front end 540. RF front end 540 may include circuitry for transmitting and receiving radio signals. For example, RF front end 540 may include receive circuitry 542 and transmit circuitry 544. In some aspects, receive circuitry 542 may be in communication with DL front end 560, which may include circuitry for receiving radio signals via antenna 335b.

[0074] In some aspects, a switch 570 may couple transmit circuitry 534 to uplink (UL) front end 572. In addition, switch 570 may couple transmit circuitry 544 to UL front end 572. UL front end 572 may include circuitry for transmitting radio signals via antenna 336. Thus, when cellular communication circuitry 330 receives instructions to transmit according to the first134904-9517-2718, v. 3RAT (e.g., as supported via modem 510), switch 570 may be switched to a first state that allows modem 510 to transmit signals according to the first RAT (e.g., via a transmit chain that includes transmit circuitry 534 and UL front end 572). Similarly, when cellular communication circuitry 330 receives instructions to transmit according to the second RAT (e.g., as supported via modem 520), switch 570 may be switched to a second state that allows modem 520 to transmit signals according to the second RAT (e.g., via a transmit chain that includes transmit circuitry 544 and UL front end 572).

[0075] As described herein, the modem 510 may include hardware and software components for implementing the above features or for determining a physical downlink shared channel scheduling resource for a user equipment device and a base station, as well as the various other techniques described herein. The processors 512 may be configured to implement part or all of the features described herein, e.g., by executing program instructions stored on a memory medium (e.g., a non-transitory computer-readable memory medium). Alternatively (or in addition), processor 512 may be configured as a programmable hardware element, such as an FPGA (Field Programmable Gate Array), or as an ASIC (Application Specific Integrated Circuit). Alternatively (or in addition) the processor 512, in conjunction with one or more of the other components 530, 532, 534, 550, 570, 572, 335 and 336 may be configured to implement part or all of the features described herein.

[0076] In addition, as described herein, processors 512 may include one or more processing elements. Thus, processors 512 may include one or more integrated circuits (ICs) that are configured to perform the functions of processors 512. In addition, each integrated circuit may include circuitry (e.g., first circuitry, second circuitry, etc.) configured to perform the functions of processors 512.

[0077] As described herein, the modem 520 may include hardware and software components for implementing the above features for determining a physical downlink shared channel scheduling resource for a user equipment device and a base station, as well as the various other techniques described herein. The processors 522 may be configured to implement part or all of the features described herein, e.g., by executing program instructions stored on a memory medium (e.g., a non-transitory computer-readable memory medium). Alternatively (or in addition), processor 522 may be configured as a programmable hardware element, such as an FPGA (Field Programmable Gate Array), or as an ASIC (Application Specific Integrated Circuit). Alternatively (or in addition) the processor 522, in conjunction with one or more of144904-9517-2718, v. 3the other components 540, 542, 544, 550, 570, 572, 335 and 336 may be configured to implement part or all of the features described herein.

[0078] In addition, as described herein, processors 522 may include one or more processing elements. Thus, processors 522 may include one or more integrated circuits (ICs) that are configured to perform the functions of processors 522. In addition, each integrated circuit may include circuitry (e.g., first circuitry, second circuitry, etc.) configured to perform the functions of processors 522.

[0079] 5G supports multi-antenna transmission, beam-forming, and simultaneous transmission from multiple geographically separates sites. Channels of different antenna ports that are relevant for a UE may differ, for example, in terms of radio channel properties. QCL antenna port may be geographically separated. Massive MIMO (mMIMO) and beamforming are used in radio access technology (RAT) cell networks such as NR and LTE. Generally, beamforming transmission uses multiple antennas to control the direction of a wave-front by appropriately weighting the magnitude and phase of individual antenna signals in an array of multiple antennas. The same signal may be sent from multiple antennas that have sufficient space between them. In any given location, the receiver (e.g., of a UE) may receive multiple copies of the same signal. Depending on the location of the receiver, the signals may be in opposite phases, destructively averaging each other out, or constructively sum up if the different copies are in the same phase, or anything in between. Analog and digital beamforming is supported in NR. In digital beamforming, a signal is pre-coded (amplitude and phase modifications) in baseband processing before RF transmission. Multiple beams (e.g., one per UE) can be formed simultaneously from the same set of antenna elements. In MU-MIMO operation, multiple TRX chains can be transmitted by a base station, one per each simultaneous MU-MIMO user. Digital beamforming improves the cell capacity as the same PRBs (frequency / time resources) can be used to transmit data simultaneously for multiple users. With analog beamforming, the signal phases of individual antenna signals are adjusted in RF domain. Analog beamforming impacts the radiation pattern and gain of the antenna array, thus improves coverage. Unlike in digital beamforming, only one beam per set of antenna elements can be formed. The antenna gain boost provided by the analog beamforming overcomes partly the impact of high pathloss in mmWave, and is therefore important for NR network operation which operates in the mmWave frequency range.

[0080] Massive multi-input multi-output (mMIMO) is a system a base station includes many antennas (e.g., 32 or more logical antenna ports per base station). For example, an154904-9517-2718, v. 3antenna array may include ‘X’ number of omni elements, with at least U wavelength spacing in between the antenna elements. The ‘X’ number of elements transmit numerous distinct streams of data via plurality of logical antenna ports, one stream for each UE. All streams are individually transmitted using the same physical resource blocks, e.g., the same time / frequency resources. The data streams do not interfere between each other because each of them has a distinct radiation pattern (through beamforming), where the signal strength in the direction of the target UE is optimized, and in the directions of the other UEs the signal strength is minimized.

[0081] 5G physical channels provide flexible communication between the 5G base stations and the UEs. 5G NR has specified the physical channels for 5G networks that can be used either for Downlink or Uplink communication. 5G NR physical channels used for uplink communication includes the physical uplink shared channel (PUSCH), the physical uplink control channel (PUCCH), and the physical random-access channel (PRACH). Uplink signals such as DM-RS, PT-RS, and SRS are also supported. 5G NR supports the simultaneous transmission on PUSCH and PUCCH. PUSCH is typically used to carry the user data and optionally, can carry uplink control information (UCI). PDSCH stands for Physical Downlink Shared Channel and is a channel used to deliver data from the base station (e.g., gNB) to the user equipment (UE) in the downlink direction. PDSCH supports high data rates and low latency for a wide range of applications and services. It uses advanced modulation and coding schemes, as well as multiple antenna techniques such as MIMO (Multiple Input Multiple Output), to maximize spectral efficiency and improve the overall performance of the network. PDSCH is also used in conjunction with other channels, such as the Physical Downlink Control Channel (PDCCH) and Physical Hybrid ARQ Indicator Channel (PHICH), to support features such as channel state information reporting, scheduling and retransmission of data packets, and HARQ (Hybrid Automatic Repeat Request) feedback. PDSCH enables the delivery of high-speed data and low-latency services to users in the downlink direction, and supports a range of advanced features and capabilities that promote efficient and reliable operation of the network.

[0082] Carrier aggregation (CA) is a technique that mobile carriers may use to increase network capacity and speeds. CA operations may include using or combining multiple carriers simultaneously to create a wider channel for data transmission and reception, which may result in increased data throughput and reduced latency. CA may be performed by splitting164904-9517-2718, v. 3and combining signals in the media access control (MAC) layer, whereas dual connectivity (DC) uses traffic splits in the Packet Data Convergence Protocol (PDCP) layer.

[0083] A UE may change physical location within a network. The network may include a plurality of cells (e.g., base stations). When UE moves from one location to another, the network may designate a different cell to be the primary cell or serving cell for the UE. In the handover procedure, a UE may send a measurement report with neighbor cell physical cell identity (PCI) and signal strength to the serving cell. The network may look at various conditions (e.g., triggering events) and if one of conditions are satisfied for a triggering event, start the handover procedure to the best target cell according to the triggering event. The target cell of the network may complete the handover procedure with the UE. The target cell may be within the same radio access technology (RAT) or a different RAT.

[0084] Generally, a network may configure a UE through radio resource control (RRC) signaling. How the network configures a UE may depend on various factors including network conditions (e.g., bandwidth, coverage, network capabilities, etc.) and UE conditions (e.g., UE capabilities, UE model, UE location, UE state, etc.). RRC signaling includes specified protocols that define how UE and network are to interact to perform connection establishment and release functions, broadcast of system information, radio bearer establishment, reconfiguration and release, RRC connection mobility procedures, paging notification and release and outer loop power control. The network may use the RRC protocols (which may be referred to as an RRC layer) to configure UE and control planes according to the network status. RRC operation includes a state machine which defines certain specific states of a UE such as RRC connected, RRC idle, etc. A UE is allocated different network resources depending on the current state of the UE.

[0085] The Access and Mobility Management Function (AMF) is a control plane network function (NF) in the 5G core network (5GC). The AMF manages access and mobility for 5G devices, including providing Registration Management, Reachability Management, Connection Management, and Mobility Management. Registration Management allows a UE to register and de-register with the 5G system. Connection Management establishes and release the control plane signaling connections between the UE and AMF. Reachability Management ensures that a UE is always reachable. Mobility Management includes network operation (e.g., signaling, tracking, and recordkeeping) to maintain knowledge where different UEs are located within the cellular network. The UE performs periodic registration174904-9517-2718, v. 3updates after it is initially registered, and may signal the network periodically with updates for the network to track each UE.

[0086] A UE or network may perform operations to enhance measurement for mobility management, in accordance with aspects of the present disclosure.

[0087] Handover may be based on L3 measurement report from UE. The network provides the L3 measurement configuration to UE. A UE performs a neighbor cell measurement and reports the measurement result to the network. When serving cell quality becomes worse, the UE starts to perform neighboring cell measurements. A UE initiates the measurement report of its neighbor when an agreed upon report condition is met (e.g., periodical report, or event triggered report). Based on the UE measurement report which the UE sends to the network, the network decides to switch to a target cell for handover.

[0088] Under L3 based mobility, L3 measurement is based on cell specific measurement performed by the UE. The UE derives a cell specific measurement result based on measuring signal strength of one or more beams of the cell (e.g., all the transmitted beams of the cell). The UE generates and sends a measurement report to the network that includes the determined cell quality, and optional beam results for the triggered cell and serving cell(s) in a measurement report (MR).

[0089] Lower layer (layer 1 and layer 2) triggered mobility (LTM) is a handover procedure that reduces the time it takes to interrupt a handover. It is a new procedure in 5G Advanced that combines the beam managing framework with the mobility framework. LTM is applicable to all frequency bands and can be used for UEs configured to use carrier aggregation. It allows a serving cell to change via L1 / L2 signaling while keeping the configuration of the upper layers intact.

[0090] Based on L3 measurement report, the network (NW) may decide one or more LTM candidate cells and provide the LTM candidate cell configuration to UE via RRC signaling. The configuration may include IDs of the candidate cells and conditions upon which to perform measurements of those cells. Based on one or more LI measurement reports from the UE, the NW changes UE serving cell by a cell switch command (e.g., through MAC CE). The candidate config ID in cell switch command refers to the configuration provided by the network to the UE through RRC in advance.

[0091] Measurement for LTM may be performed by a UE via two different procedure, an L3 measurement procedure and LI measurement procedure. Regarding L3 measurement procedure, UE may use existing L3 measurement procedure for LTM. The UE performs L3184904-9517-2718, v. 3measurement procedure for LTM to help the network decide which is a preferred LTM candidate, and provides the LTM candidate configuration to the UE with the preferred LTM candidate. Regarding LI measurement procedure, the UE performs LI measurement upon receiving the LTM candidate configuration. Based on the LI measurement report of the candidate cell, the network makes the LTM decision and sends to the UE, an LTM cell switch command (e.g., via MAC CE) to trigger UE to perform cell switch. As such, existing measurement for UE mobility can be summarized as follows: mobility mechanisms are based on L3 measurement. The UE performs L3 measurement, and the network uses the result for handover (HO) and / or conditional handover (CHO), and a combination of L3 measurement and LI measurement is used for LTM mobility. Further, L3 measurement procedure is designed based on cell specific measurement result - both the reported measurement results and the trigger conditions are based on cell specific results rather than beam specific results (e.g., a strength of a single beam).

[0092] Channel State Information (CSI) is a set of known properties of a communication link in wireless communications. CSI is used in 5G to measure the characteristics of a radio channel, which helps ensure that the correct modulation, code rate, and beam forming are used. This information describes how a signal propagates from the transmitter to the receiver and represents the combined effect of, for example, scattering, fading, and power decay with distance. Measuring characteristics of a channel in such a manner may be referred to as channel estimation. CSI makes it possible for the network to adapt transmissions to current channel conditions, which helps provide a reliable communication with high data rates in multiantenna systems. CSI is estimated at the receiver and typically quantized and reported to the transmitter, although reverse-link estimation is possible in time-division duplex (TDD) systems. Therefore, the transmitter and receiver can have different CSI. The CSI at the transmitter and the CSI at the receiver are sometimes referred to as CSIT and CSIR, respectively.

[0093] Within the 3GPP Technical Specification Group Radio Access Network (TSG RAN), RAN WG4 (RAN4) is responsible for the following aspects. The group works on the Radio Frequency (RF) aspects of UTRAN / E-UTRAN / NR. RAN4 performs simulations and derives minimum requirements for transmission and reception parameters, Radio Resource Management (RRM), and for channel demodulation and Channel State Information (CSI) reporting.194904-9517-2718, v. 3

[0094] Once these requirements are set, the group defines the test procedures that will be used to verify them for BS, Repeater, Relay and Integrated Access Backhaul (IAB).

[0095] Requirements for other radio elements, like Repeaters, Relay, and IAB are specified in RAN4 as well. In addition, OTA requirements and testing methods for both BS and UE are defined in RAN4.

[0096] A 2-sided ML model may comprise an encoder and decoder that performs CSI compression. The encoder may be referred to as a test encoder and the decoder may be referred to as a test decoder. The encoder represents the role of the UE in the network, and the decoder represents the role of the gNB in the network. There are four options for generating a 2-sided ML model. In option 1, a DUT vendor provides the test decoder. In option 2, a decoder vendor serves as a source of the decoder.

[0097] Under option 3, the test decoder specifications are simply agreed upon, specified (e.g., in RAN4), and followed universally.

[0098] Under option 4, each test equipment (TE) vendor provides their own decoder, which is developed based on an agreed upon specification (e.g., RAN4).

[0099] Table 1 below contains a set of parameters which may be implemented, for example, under option 3. Table 1 indicates how the test decoder could be implemented (e.g., trained) based on a full set of RAN4 specifications. Other parameters that are not yet listed might also be important.Table 1204904-9517-2718, v. 3214904-9517-2718, v. 3

[0100] The objective of a neural network is to serve as a final model that performs well not only on the training dataset but also on new data, enabling accurate predictions. The ability to perform well on previously unobserved inputs is called generalization. A related concept to generalization is overfitting. Overfitting is the condition where a model learns and adapts to a training dataset too well. The overfit model performs well on the training dataset but may performs poorly or less well on input data that is different from the training dataset. An effective way to reduce overfitting and improve generalization is by constraining the model complexity, which may be referred to as regularization.

[0101] Two ways to approach to avoid overfitting a model include (1) to reduce overfitting by training the network on more examples, and (2) to reduce overfitting by changing the complexity of the network.

[0102] A benefit of very deep neural networks is that their performance continues to improve as they are fed larger and larger datasets. A model with a near-infinite number of examples will eventually plateau in terms of what the capacity of the network is capable of learning.

[0103] A model can overfit a training dataset because it has sufficient capacity to do so. Reducing the capacity of the model reduces the likelihood of the model overfitting the training dataset, to a point where it no longer overfits. Techniques that can be employed to avoid overfitting may include activity regularization (penalize the model during training base on the224904-9517-2718, v. 3magnitude of the activations), weight constraint (constrain the magnitude of weights to be within a range or below a limit), dropout (probabilistically remove inputs during training), noise (add statistical noise to inputs during training), and early stopping (monitor model performance on a validation set and stop training when performance degrades).

[0104] In an aspect, when specifying the test decoder for Option 3, the specification (e.g., RAN4) considers supplementing the list of parameters to fully include regularization techniques, such as weight constraints, dropout layers and their associated probabilities, early stopping criteria, noise injection, and other relevant factors.

[0105] In an embodiment, training a two-sided model or fully specified test decoder (with tuned weights and biases for each node) may be performed according to specification such as, for example: the pair of reference encoder / test decoder can be trained jointly with the specified training data set; the parameters used to generate the training data (channel model and CSI configuration), and / or details regarding the format of input CSI data to the reference encoder, such as, for example, whether the inputs to the encoder comprise raw channels or the eigenvectors, or specifying a post processing procedure for the channel estimation output.

[0106] Once the test decoder has been implemented (e.g., trained and tested) based on the specifications and agreements, each participating company can provide their own designed test decoder.

[0107] The channel model may be specified as part of the training data set. During the actual testing procedure, the DUT (e.g., UE encoder) will be tested along with the test decoder. However, the input to the test decoder may be trained with the output of a test encoder (different from that of the DUT) and that test encoder will be trained based on the agreed- upon dataset. As such, there remains a risk that the behavior of test decoder may not be accurate when it interacts with the encoder on the DUT.

[0108] Option 3 (e.g., fully specifying a test decoder in RAN4) may present various issues to be addressed.

[0109] For instance, the source of the original training data should be determined. For example, if results are captured only based on 3GPP statistical channel models, then there is less certainty on how well these results represent real-world measurements. Hence, there is little guidance as to what training data should be captured and used for testing to achieve reliable performance across a large variety of deployment scenarios, which reflect real world deployment of such models.234904-9517-2718, v. 3

[0110] Each NW vendor and each UE vendor may have some specific variant implementations. If a test decoder is not trained for different UE encoder implementations, option 3 may not be feasible given that even a well-trained UE encoder could encounter unpredictable behavior from a test decoder if that test decoder has not been trained to interact with the reference encoder. A test decoder cannot realistically produce meaningful answers for all expected device under test (DUT) UE. A well-trained UE with ‘encoder A’ that has been trained with a ‘dataset A’ may fail test with ‘test decoder B’ which has been trained with a different ‘dataset B’, especially if that ‘test decoder B’ is not able to generalize. UE DUT could have a different backbone network, and / or a different backbone network from a reference encoder B that ‘teste encoder B’ was trained with. The test decoder should also capture characteristic of NW vendors, because the testing (e.g., test encoder B and other TE) should reflect real conditions in the actual deployment. Otherwise, testing of a UE DUT could pass at TE vendor but fail during real deployment.[OHl] The input characteristics to the test decoder could be different if the testing dataset is different than the standardized training dataset.

[0112] An important consideration may be whether the information provided in the specification and the testing dataset used to train the test decoder are adequate to ensure reliable performance across all deployment scenarios. For alignment with real-world deployment, it may be important that network vendors possess sufficient information to expect the gNB decoders to perform the same or within a threshold similarity to behavior of the test decoder. Achieving this alignment with standardized decoders may pose challenges, as it may make a high degree of similarity between the gNB decoder and the test decoders important.

[0113] Defining one standardized training dataset (configuration / scenario, channel model) may be restrictive. To cover more conditions, multiple standardized datasets and multiple test decoders could be used. That could complicate the testing procedures.

[0114] As such, Option 3 may present numerous significant issues to be addressed in order to be a feasible option for a testing framework for AI / ML models for NR Air Interface.

[0115] In an aspect of the present disclosure, to assess the feasibility of option 3, RAN4 should address several key considerations. These include potential mismatches between the standardized training dataset and the testing dataset, as well as discrepancies between the Device Under Test (DUT) encoder and the reference decoder (e.g., a decoder trained according to the specification of RAN4). Additionally, such an approach should consider the244904-9517-2718, v. 3challenges related to potential differences between testing conditions and real-world deployment.

[0116] When defining a static scenario or configuration, such as through the definition of a related testing dataset based on channel models outlined in TR 38.901, datasets derived from TR 38.901, such as UMa, UMi, CDL channels, and the 'legacy approach,' serve as initial reference points. However, it is important to note that limiting testing to the legacy approach, such as TDL channels, presents challenges in asserting that performance test outcomes accurately reflect the dependability and effectiveness of the model being tested.

[0117] In an aspect, the CDL channel and / or UMa channel can be prioritized as the baseline for static scenario / configuration testing. The simplicity of the TDL channel modeling method may make it inadequate for accurately mimicking the complexities of data-driven AI / ML algorithms. Even if a model performs well on the TDL channel during testing, it is challenging to ascertain its effectiveness on Urban Macrocell (UMa) channels, Clustered Delay Line (CDL) channels, and especially complex field channels, particularly in cases involving CSI compression. Therefore, it is beneficial to prioritize the CDL or UMa channel models from TR 38.901 as the baseline for two-model training or testing (e.g., static scenario / configuration testing).

[0118] As mentioned, under option 4, each test equipment (TE) vendor provides their own decoder, which is developed based on an agreed upon specification (e.g., RAN4). Based on the previous discussions of option 4 with a two-sided model, various issues have been discussed.

[0119] Option 1 - Capturing a dataset {encoder input, encoder output, decoder output} in RAN4 specification, and the test decoder implementations are verified against this dataset. Two sub-options for dataset generation are listed below: Option 4a- 1 : The dataset is generated by one agreed reference encoder / decoder pair (for dataset generation purpose); and Option 4a-2: The dataset is generated by the encoder / decoder pairs designed by the contributing companies based on the agreed common assumptions.

[0120] Option 2: Capturing the encoder in the agreed reference encoder / decoder pair (for test decoder verification purpose) in RAN4 specification, and the test decoder implementations are verified against this reference encoder.

[0121] Option 3: discuss which parameters / assumptions should be different / not applicable compared to Option 3 in Issue 4-3.254904-9517-2718, v. 3

[0122] Option 4: Model structure is not specified in RAN4. Training dataset is specified, where each training sample consists of both the raw channel matric / precoding matrix and the bit stream forwarded to the test decoder.

[0123] Option 5: Model structure is specified in RAN4. Training dataset is not specified for verifying the encoder at DUT. The test decoder developed by TE vendor should be verified.

[0124] In aspects of the present disclosure, another option that addresses the issues described above is disclosed. Regarding option 4, for decoder source for two-sided model test, a number of considerations may be relevant to providing robust TE.

[0125] For example, the TE vendor should be able to develop the decoder based on the specifications. Test repeatability should be ensured (variation among TE vendor implementations should be bound). Other vendors should also be able to develop such a decoder and which can deliver similar performance. The parameters of importance should be specified, and interoperability should be ensured based on those parameters. Candidate parameters / conditions that may be considered for defining a test decoder may include training data set for TE decoder training, model structure (activation function is included in the model structure), performance parameters for the TE decoder (e.g., cosine similarity, loss function, etc.), maximum FLOPs allowed for the test decoder, maximum number / size of model parameters, and / or compression ratio of decoder (output size / input size).

[0126] Under Option 4, it is desirable for a single decoder that is implemented by each TE vendor to be sufficient to perform a single test for any DUT. A TE vendor should be able to implement the test decoder for Option 4 without any involvement from another party. If this assumption is infeasible, another option in which TE vendors should collaborate with DUT / infra vendors to implement the decoder could be considered.

[0127] Further clarification and analysis of the four options of test decoder are considered. It is assumed that for Option 4 the TE vendors can implement the decoder just based on the specifications, with no other parties involved. For Option 4, the decoder is developed by each TE vendor and test repeatability should be ensured.

[0128] In order to bound the AI / ML output variation within certain range, some parameters of the test decoder should be specified. That is, sufficient information should be captured in the specification of the decoders to enable reliable and repeatable testing. A DUT UE with encoder that can pass on one TE (with test decoder) will be able to pass on another TE (with different test decoder) and provide similar performance with a gNB that has a decoder trained in a similar manner. Adopting this practice with option 4 may allow each TE vendor to264904-9517-2718, v. 3independently design their own decoder according to the specification, rather than a decoder being passed between vendors.

[0129] A number of issues may be addressed to ensure the feasibility of option 4. Such as, for example, determining specific information is to be included in the decoder specification to ensure that decoders are trained in a manner that enables repeatable testing and ensures reliable performance across all deployment scenarios, determining the source of training data, and whether it is feasible for the specified training dataset to bound the variations in TE implementations and performance, reduce model mismatches and guarantee generalization in real deployment. Aspects of the present disclosure address such issues for option 4.

[0130] For the decoder specification it is important to consider several key metrics commonly used in AI / ML algorithms to gauge processing capability. These metrics include floating-point operations per seconds (FLOPs) and the quantity / size of parameters, which evaluate computational complexity and the memory used by AI / ML models, respectively. Additionally, it may be important to explicitly specify specific parameters like model type and structure and activation functions.

[0131] It is observed that, to bound the variations in the TE implementations for the encoder / decoder for option 4, some parameters of the module structure should be specified by RAN4.

[0132] In an aspect, the following parameters may be specified for test decoder for option 4 by RAN4: Model type such as the backbone of the model, examples of which may include convolutional neural network (CNN), Transformer, multi-layer perceptron (MLP), or other architecture type; Model structure such as, for example, depth, width, etc.; activation functions of the model; maximum FLOPs allowed for the test decoder; and / or maximum number / size of parameters.

[0133] Option 4 for test decoder dictates that the TE should be able to implement the decoder based on some specifications. For implementing the test decoder, an AI / ML model architecture, including its layers, activations, normalization methods, etc., as well as their corresponding numerical weights, may be fully defined. To obtain each of the numerical weights, a training dataset is used to train the model and tune these weights to reduce error between the output of the model and the desired behavior.

[0134] For the source of the training dataset for the RAN4-specified test decoder there is a consideration and desire for constructing a standardized data set or specifying some parameters to help develop such a dataset. If a data set used for training the test decoder is274904-9517-2718, v. 3specified, vendors (both TE and infra-vendors) can develop a test decoder which can deliver similar performance, reduce the output variation and ensure repeatability, this decision holds significant weight as the dataset substantially influences the performance of AI / ML models, typically tailored to specific scenarios. However, should RAN4 opt to define dataset(s) for the specified decoder, multiple datasets for varying scenarios might emerge. Yet, a pertinent issue arises: new AI / ML algorithms or scenarios may struggle to meet performance expectations if the dataset(s) become outdated.

[0135] RANI relies on results derived from 3 GPP statistical channel models, lacking conclusions regarding how performance can be assured in real-world measurements. As a result, there is uncertainty regarding the important metrics or methodologies to be used to achieve consistent performance across various deployment scenarios.

[0136] It is observed that there are multiple issues with the adaptation of a standardized training dataset. The issues range from providing multiple training datasets, to dataset obsolescence and generalizability concerns in real deployment.

[0137] For CSI compression using two-sided model use case, AI / ML model training collaborations may include different types. Type 1 refers to joint training of the two-sided model at a single side / entity, e.g., UE-sided or Network- sided. Type 2 refers to joint training of the two-sided model at network side and UE side, respectively. Type 3 refers to separate training at network side and UE side, where the UE-side CSI generation part and the NW-side CSI reconstruction part are trained by UE side and network side, respectively.

[0138] FIG. 6 shows an example of training a test decoder with type 3 training. With type 3 training a test decoder 604 could be trained with a dataset consisting of CSI target and feedback CSI data starting with UE side training first. During operation 1 of UE-first type 3 training, the first vendor initiates by training an autoencoder 602.

[0139] Subsequently, at operation 2, the first vendor shares a labeled dataset {V, c} with a second vendor. This labeled dataset may comprise training dataset pairs of compressed (e.g., quantized akin to a CSI report) CSI data ‘c’ and corresponding target CSI data ‘V’ or ‘ V-haf , which may be sourced from the output of first decoder 606.

[0140] At operation 3, the second vendor trains a decoder 604 with the objective to align decoder 604 with the encoder 602 developed by the first vendor, by adjusting the numeric weights of decoder 604 through training to minimize the loss between the test data ‘V’ and output of the decoder 604 shown as ‘V-tilda’.284904-9517-2718, v. 3

[0141] If the training dataset is specified as being generated at step 1 the following issues may be considered:

[0142] In the TR, RANI notes a decline in performance when comparing the originally trained model with the type-3 trained model. The implications of this decline for the dependable reproducibility of a test, as well as its adequacy in ensuring reliable performance, remain uncertain.

[0143] The observed performance loss appears to be more pronounced when the backbones (such as general architectures like fully connected, CNN, Transformer, variations of autoencoders) differ compared to when they remain the same. Consequently, there is ambiguity regarding the extent of flexibility available for training different decoders, both from the TE vendor and the infrastructure vendor perspectives. This is due to the mismatch between the generated training data and their subsequent use to train the test decoders.

[0144] The source of the labeled training data {V, c} could originate from different UE encoders (since it is UE first training). RANI observes even more significant deterioration, and in some cases, catastrophic deterioration, when the dataset used to train a decoder originates from multiple trained encoders.

[0145] It is unclear what the source of the training data would be for this training scenario, and whether this should be specified in RAN 1 or RAN 4 to provide the training data source.

[0146] Therefore, there is currently no definitive answer regarding how or what elements should be captured to attain reliable performance across all deployment scenarios. As such, it is observed that there are performance issues if the training of test decoder 604 is done with type 3 training. The performance loss of the test decoder 604 depends on the discrepancy between the backbones of encoder 602 and decoder 604, and whether the training dataset consists of a diverse dataset from multiple trained encoders. There is currently a lack of definition of training elements that should be specified to attain reliable performance across all deployment scenarios.

[0147] Helping to address those issues, aspects of the present disclosure implement a jointly trained test decoder with a TE implemented UE encoder (proxy Encoder). This means that both the encoder and decoder will be implemented using TE type 1 training - joint training of the two-sided model at a single entity (UE or network).

[0148] FIG. 7 shows an example system for implementing a test decoder, in accordance with an embodiment. In FIG. 7, a framework for training test decoder for option 4 and UE encoder testing.294904-9517-2718, v. 3

[0149] In an aspect, test decoder implementation may employ type 1 training to jointly train the test decoder with a TE implemented UE encoder, which may be referred to as a proxy encoder. Joint training of the two-sided model may be performed at a single entity (UE or network). This may better support the Option 4 framework, where each TE vendor implements their own test encoder.

[0150] A system 700 may comprise one or more computing devices, configured to perform operations to jointly train the two-sided model 704. The system 700 may obtain training data 702 that is associated with channel state information (CSI) and a cellular network (e.g., a 5G NR cellular network). System 700 may input the training data 702 to the two-sided machine learning (ML) model 704. The two-sided ML model 704 may comprise a 706, which may be referred to as a proxy encoder, and a 708 which may be referred to as a test equipment (TE) decoder.

[0151] In an embodiment, each side of the two-sided ML model may comprise a backbone such as, for example, a CNN, a transformer, an MLP, or other backbone. In an embodiment, ML model 704 may represent a single ML model, or a plurality of ML models.

[0152] System 700 may compress, by the encoder, the channel state information (CSI) resulting in compressed CSI data 710. System 700 may decode, by the TE decoder 708, the compressed CSI data 710, to obtain second CSI data V-hat. During training, the respective weights between n

[0153] The source of training data 702 could be designed to be more flexible than a standardized dataset by employing a cooperative approach where diverse datasets are being collected to provide a mixture of training data that could aid the generalization ability of the models. Training data processor training data processor 728 may comprise processing logic (e.g., database, server, algorithms, etc.) to process and store a database of measurement. This training data may be sourced at the TE side, where data from statistical models 718, NW infrastructure vendors 714, UE vendors 712, and TE vendors 716 are aggregated and utilized to train both the proxy encoder 706 and decoder 708 within the TE system 700.

[0154] The collected data from 712, 714, 716, 718 may be combined to produce synthetic data 720 which are artificially generated data that is designed to mimic real-world data. Synthetic data 720 may be generated using statistical distributions or generative models 722 that replicate the patterns, characteristics, and relationships found in real-world data, statistical distributions or generative models 722 may comprise a statistical model (e.g., a304904-9517-2718, v. 3parametric model), a machine learning model, such as, for example, a neural network, or other algorithm or algorithms.

[0155] In an aspect, implementing the test decoder 708 in Option 4 and for its training procedure may comprise using synthetic data 720 to train the two-sided model 704. Synthetic data 720 is generated by statistical distributions or generative models 722, to ensure that the resulting training data 702 is sufficiently diverse, yet still captures real network characteristics and diverse scenarios.

[0156] In an embodiment, at operation 724, a subsequent type 3 training may be employed by leveraging the CSI feedback data ('c') generated from the proxy encoder 706 to aid training of a DUT encoder 726 which represents UE-side compressing behavior. After the DUT UE 726 encoder is trained, the testing procedure can be employed by deploying the trained UE encoder 726 and trained test decoder 708 as it is shown in FIG. 7.

[0157] In an embodiment, each subset of training data may comprise conditions 734 that are associated with UE capability, and additional conditions 736 that are not associated with UE capability, or other aspects not satisfied. Each subset of training data 730 may be generated based on different conditions 736 and / or different additional conditions 736, to create a desired collection training data 702 of scenarios and conditions.

[0158] In an embodiment, multiple databases (e.g., managed at training data processor training data processor 728) can be built depending on different scenarios / configurations, conditions and additional conditions. In an embodiment, each UE may comprise or access trained AI / ML models 708 and / or 726, where each model is associated with a particular training dataset that reflects specific conditions / scenarios and additional conditions, as described in FIG. 8.

[0159] FIG. 8 illustrations a diagram for identification, training, or testing of Al-based models, in accordance with an embodiment. The diagram shows functionality identification and NW-side selection of Al-based models for testing, in accordance with an embodiment.

[0160] As described, multiple databases (e.g., managed at training data processor training data processor 728) can be built depending on test data 802, and different scenarios / configurations, conditions and additional conditions such as training data 804. A UE 806 may comprise or access trained AI / ML models 808 where each model (e.g., an encoder or decoder) is associated with a particular training dataset that reflects specific conditions and / or scenarios. Models 1-N are shown. Some of the conditions 810 can be signaled based on additional information 804 along with the test data 802. Assistance314904-9517-2718, v. 3information in 804 may comprise metadata (e.g., an identifier, a number, a name, a summary) that may be used by the UE 806 to select the appropriate model. Assistance information 804 may comprise information and / or an indication describing a NW side condition. The UE 806 may use this assistance information 804 to select a model 812, which may be trained with the same conditions or additional conditions as that of the configured test data 802.

[0161] Assuming that a UE 806 may access or comprise multiple trained models associated with different datasets that reflect various scenarios or conditions, multiple decoders (e.g., TE decoders) may be trained with a specific dataset.

[0162] In an aspect, multiple test TE decoders may be trained (e.g., according to FIG. 7) with a specific dataset. Each dataset may comprise data that is associated with distinct scenario conditions as well as additional conditions.

[0163] In an aspect, to support the various AI / ML functionalities identified with conditions / additional conditions, a training set ID 732 can link each set 730 of training data with specific conditions to the test decoder. During testing, the system 700 may use this ID to deploy the appropriate TE decoder (trained with that data set). This may be performed with an Option 4 framework where each TE implements their own TE decoder.

[0164] Such a framework may be used when the testing framework uses multiple decoders, which may be the case in FIG. 7 in some embodiments. Different sets 730 of synthetic data 720 can be generated, each accounting for the additional conditions (e.g., network conditions). A unique training ID can serve as a label to the training data that a given model (e.g., a TE decoder or encoder) was trained with. During testing phase the training ID system 700 may select the appropriate test decoder (corresponding to the training ID) to test a DUT UE encoder.

[0165] In an aspect, in developing the TE decoder 708 for option 4 and to accommodate the various AI / ML functionalities associated with different conditions and additional factors, a mixed dataset is used for training the TE decoder 708. This may help alleviate the burden on the testing framework. In an embodiment, a dataset 730 may comprise a mixed dataset that is associated with multiple scenarios, conditions, and additional factors that together provide a variety of NW scenarios, conditions, and / or environments. Depending on the extent of performance degradation, testing complexity could be reduced by employing a single test decoder trained with such a mixed dataset.

[0166] By sourcing training data from the UE / gNB at the Test Environment (TE), and by using similar data to train the NW decoder for real deployment, system 700 may mitigate324904-9517-2718, v. 3mismatches between TE and real-world deployment. For example, the TE decoder 708 trained through system 700 may more closely resemble a gNB with decoder that is deployed in a cellular network. When a UE encoder passes test with the TE decoder 708, then it can also more reliably interact with the real gNB. In an embodiment, system 700 and training data processor 728 may continuously update training data 702 by processing contributions from different sources such as UE vendors 712, NW vendors 714, TE vendors 716, and / or statistical channels 718, thereby providing an adaptive framework where different TE vendors may more reliably test UE behavior. This would enable both UE encoders and test decoders to evolve and continuously learn, thereby mitigating potential issues associated with dataset obsolescence.

[0167] As such, an adaptive strategy that establishes an agreed upon dataset and implements a collaborative approach with data contributions received from multiple entities would be beneficial for training and implementing reliable TE (e.g., TE decoder 708), especially in the framework of Option 4.

[0168] Referring back to FIG. 8, once multiple TE decoders are trained based on respective datasets, they can be used to test UE 806. For a given test with test data 802, the DUT UE 726 may select an appropriate test based on assistance information 804 (e.g., an ID associated with test data 802). For this test, the TE system (which may be the same or different from system 700) may select the TE decoder that also corresponds to test data 802 and / or assistance information 804, such as, for example, by ID.

[0169] FIG. 9 illustrates an example method 900 for implementing test equipment (TE), in accordance with an embodiment. Although the example method depicts a sequence of operations, the sequence may be altered without departing from the scope of the present disclosure. For example, some of the operations depicted may be performed in parallel or in a different sequence that does not materially affect the function of the method. In other examples, different components of an example device or system that implements the method may perform functions at substantially the same time or in a specific sequence.

[0170] Method 900 may be performed by processing logic of one or more computing devices. Processing logic may include a combination of hardware (e.g., passive or active electronic components, programmable logic, a processor, a baseband processor, a transmitter, a receiver, an antenna, etc.) and software (e.g., machine-executable instructions stored in computer-readable memory).334904-9517-2718, v. 3

[0171] Aspects associated with method 900 may correspond to aspects described with respect to other figures, to train a two-sided machine learning model, such as described with respect to FIG. 6, FIG. 7, or FIG. 8.

[0172] At operation 902, the method comprises obtaining training data associated with channel state information (CSI) and a cellular network. In an embodiment, obtaining the training data comprises obtaining portions of the training data from a plurality of data sources (e.g., 712, 714, 716, 718), and generating synthetic data (e.g., 720) from the portions of the training data. In an embodiment, generating the synthetic data comprises applying a generative model or a statistical model (e.g., statistical distributions or generative models 722) to the portions of the training data to generate the synthetic data.

[0173] In an embodiment, the method prioritizes with the training data, a clustered delay line (CDL) channel or an Urban Macrocell (Uma) channel. For example, the training data (e.g., synthetic data) may be generated to comprise a disproportionate amount of data associated with CDL or Uma, or use either as a default channel condition. In such a manner, the trained encoder and decoder may reliably encode and decode CSI data under real world CLD or Uma channel conditions.

[0174] At operation 904, the method comprises inputting the training data to one or more two-sided machine learning (ML) models, each comprising an encoder and a decoder. The output of the encoder may be coupled to the input of the decoder. Each of the encoder and decoder may comprise a neural network with layers of nodes, and weighted interconnections between nodes of different layers.

[0175] At operation 906, the method comprises compressing, by the encoder the channel state information (CSI) resulting in compressed CSI data. The method may provide the input CSI training data as raw channel data or eigenvectors of a channel. The encoder may compress CSI data into a compressed bit stream.

[0176] At operation 908, the method comprises decoding, by the decoder the compressed CSI data, to obtain second CSI data. The second CSI data is a reconstructions of the training CSI data. The decoder is trained (e.g., weights adjusted) to minimize loss between the training data CSI data (fed to the encoder) and the output of the decoder (the second CSI data).

[0177] In an embodiment, the method further comprises enforcing at least one of: weight constraints, dropout layers, early stopping criteria, or noise injection. For example, to train the two-sided model, the method may monitor and set weight constraints on weights of the344904-9517-2718, v. 3ML model, by checking whether each weight exceeds a threshold. If so, those weights are rescaled so that they no longer exceed the threshold. Additionally, or alternatively, the method may remove neural network nodes in the ML model and connections of each dropped node, creating a new network architecture out of the parent network. The nodes may be randomly dropped based on applying a dropout probability of p (e.g., a ration or percentage) during forward and / or backward propagation of the ML model. Early stopping criteria may refer to one or more test conditions that, when satisfied, cause the method to stop training the ML model. Noise injection may include adding random or statistical noise (e.g., changing the test data in an insignificant way). Each of these techniques may be performed by itself or in combination with another to reduce risk of overfitting of the two- sided ML model, thereby providing a more robust encoder and decoder.

[0178] In an embodiment, the one or more two-sided ML models comprise a plurality of ML models, each trained with a respective one of a plurality of subsets of the training data. For example, each of the plurality of the subsets of training data (e.g., training data sets 730) may be associated with distinct scenarios comprising different channel conditions, different network conditions, or different combinations of a channel condition and network condition. In an embodiment, the different network conditions comprises at least one of an indoor propagation environment, an outdoor propagation environment, a location, or a distance. In an embodiment, the different channel conditions comprises at least one of a different doppler speed, a different carrier frequency, a different channel. In an embodiment, the plurality of subsets of the training data are each associated with a training set identifier (e.g., ID 732) that is associated with a respective one of the plurality of ML models trained with that subset of the training data. Additionally, or alternatively, one of the plurality of ML models may be trained by one of the subsets of the training data that is associated with a mix of the distinct scenarios, which may provide a more robust or general decoder that is less prone to overfitting.

[0179] In an embodiment, the training data comprises multiple datasets, each comprising a range of different doppler speeds, different carrier frequencies, different propagation environments (indoor, outdoor).

[0180] In an embodiment, the decoder satisfies a specification defining a model type, a model structure, an activation function, a maximum floating-point operations per seconds (FLOPs), or a maximum number or size of parameters. For example, the method may comprise performing a check of each of these conditions, or monitoring these conditions354904-9517-2718, v. 3during testing and detecting if any of the conditions are violated. If so, the method may comprise logging the violation, generating an alert, halting testing or training, etc.

[0181] In an embodiment, the method comprises using an output of the encoder (e.g., the compressed CSI data) of the two-sided model to train an encoder of a UE. Similarly, the method may comprises using the output of the decoder of the two-sided model to train a decoder of a gNB.

[0182] In an embodiment, the trained decoder of the two-sided ML model may be used as test equipment (TE) to test a DUT UE. In an embodiment, depending on the test data to test the DUT UE, the method may comprise selecting a corresponding trained decoder. For example, the test data may comprise an identifier or information (e.g., assistance information 804) to select the decoder that was trained based on that test data.

[0183] Portions of what was described above may be implemented with logic circuitry such as a dedicated logic circuit or with a microcontroller or other form of processing core that executes program code instructions. Thus, processes taught by the discussion above may be performed with program code such as machine-executable instructions that cause a machine that executes these instructions to perform certain functions. In this context, a “machine” may be a machine that converts intermediate form (or “abstract”) instructions into processor specific instructions (e.g., an abstract execution environment such as a “virtual machine” (e.g., a Java Virtual Machine), an interpreter, a Common Language Runtime, a high-level language virtual machine, etc.), and / or, electronic circuitry disposed on a semiconductor chip (e.g., “logic circuitry” implemented with transistors) designed to execute instructions such as a general-purpose processor and / or a special-purpose processor. Processes taught by the discussion above may also be performed by (in the alternative to a machine or in combination with a machine) electronic circuitry designed to perform the processes (or a portion thereof) without the execution of program code.

[0184] The present invention also relates to an apparatus for performing the operations described herein. This apparatus may be specially constructed for the required purpose, or it may comprise a general-purpose computer selectively activated or reconfigured by a computer program stored in the computer. Such a computer program may be stored in a computer readable storage medium, such as, but is not limited to, any type of disk including floppy disks, optical disks, CD-ROMs, and magnetic-optical disks, read-only memories (ROMs), RAMs, EPROMs, EEPROMs, magnetic or optical cards, or any type of media suitable for storing electronic instructions, and each coupled to a computer system bus.364904-9517-2718, v. 3

[0185] A machine-readable medium includes any mechanism for storing or transmitting information in a form readable by a machine (e.g., a computer). For example, a machine- readable medium includes read only memory (“ROM”); random access memory (“RAM”); magnetic disk storage media; optical storage media; flash memory devices; etc.

[0186] A baseband processor (also known as baseband radio processor, BP, or BBP) is a device (a chip or part of a chip) in a network interface that manages radio functions, such as communicating (e.g., TX and RX) over an antenna.

[0187] An article of manufacture may be used to store program code. An article of manufacture that stores program code may be embodied as, but is not limited to, one or more memories (e.g., one or more flash memories, random access memories (static, dynamic, or other)), optical disks, CD-ROMs, DVD ROMs, EPROMs, EEPROMs, magnetic or optical cards or other type of machine-readable media suitable for storing electronic instructions. Program code may also be downloaded from a remote computer (e.g., a server) to a requesting computer (e.g., a client) by way of data signals embodied in a propagation medium (e.g., via a communication link (e.g., a network connection)).

[0188] The preceding detailed descriptions are presented in terms of algorithms and symbolic representations of operations on data bits within a computer memory. These algorithmic descriptions and representations are the tools used by those skilled in the data processing arts to most effectively convey the substance of their work to others skilled in the art. An algorithm is here, and generally, conceived to be a self-consistent sequence of operations leading to a desired result. The operations are those requiring physical manipulations of physical quantities. Usually, though not necessarily, these quantities take the form of electrical or magnetic signals capable of being stored, transferred, combined, compared, and otherwise manipulated. It has proven convenient at times, principally for reasons of common usage, to refer to these signals as bits, values, elements, symbols, characters, terms, numbers, or the like.

[0189] It should be kept in mind, however, that all of these and similar terms are to be associated with the appropriate physical quantities and are merely convenient labels applied to these quantities. Unless specifically stated otherwise as apparent from the above discussion, it is appreciated that throughout the description, discussions utilizing terms such as “transmitting”, “sending”, “selecting,” “determining,” “receiving,” “forming,” “grouping,” “aggregating,” “generating,” “removing,” or the like, refer to the action and processes of a computer system, or similar electronic computing device, that manipulates and transforms374904-9517-2718, v. 3data represented as physical (electronic) quantities within the computer system's registers and memories into other data similarly represented as physical quantities within the computer system memories or registers or other such information storage, transmission or display devices.

[0190] The processes and displays presented herein are not inherently related to any particular computer or other apparatus. Various general-purpose systems may be used with programs in accordance with the teachings herein, or it may prove convenient to construct a more specialized apparatus to perform the operations described. The required structure for a variety of these systems will be evident from the description below. In addition, the present invention is not described with reference to any particular programming language. It will be appreciated that a variety of programming languages may be used to implement the teachings of the invention as described herein.

[0191] It is well understood that the use of personally identifiable information should follow privacy policies and practices that are generally recognized as meeting or exceeding industry or governmental requirements for maintaining the privacy of users. In particular, personally identifiable information data should be managed and handled so as to minimize risks of unintentional or unauthorized access or use, and the nature of authorized use should be clearly indicated to users.

[0192] The foregoing discussion merely describes some exemplary aspects of the present invention. One skilled in the art will readily recognize from such discussion, the accompanying drawings and the claims that various modifications can be made without departing from the spirit and scope of the invention.384904-9517-2718, v. 3

Claims

CLAIMSWhat is claimed is:

1. A system comprising one or more computing devices, configured to perform operations comprising: obtaining training data associated with channel state information (CSI) and a cellular network; inputting the training data to one or more two-sided machine learning (ML) models, each comprising an encoder and a decoder; compressing, by the encoder, the CSI, resulting in compressed CSI data; and decoding, by the decoder, the compressed CSI data, to obtain second CSI data.

2. The system of claim 1, wherein the operations further comprise enforcing at least one of: weight constraints, dropout layers, early stopping criteria, or noise injection.

3. The system of claim 1, wherein the training data comprises data associated with a clustered delay line (CDL) channel or an Urban Macrocell (Uma) channel.

4. The system of claim 1, wherein obtaining the training data comprises: obtaining portions of the training data from a plurality of data sources, and generating synthetic data from the portions of the training data, wherein the training data comprises a combination of the portions of the training data and the synthetic data.

5. The system of claim 4, wherein generating the synthetic data comprises applying a generative model or a statistical model to the portions of the training data to generate the synthetic data.

6. The system of claim 1, wherein the one or more two-sided ML models comprise a plurality of ML models, each trained with a respective one of a plurality of subsets of the training data.

7. The system of claim 6, wherein each of the plurality of the subsets of training data are associated with distinct scenarios comprising different channel conditions, different network conditions, or different combinations of a channel condition and network condition.394904-9517-2718, v.

38. The system of claim 7, wherein the different network conditions comprises at least one of: an indoor propagation environment, an outdoor propagation environment, a location, or a distance.

9. The system of claim 7, wherein the different channel conditions comprises at least one of: a different doppler speed, a different carrier frequency, a different channel.

10. The system of claim 6, wherein the plurality of subsets of the training data are each associated with a training set identifier (ID) that is associated with a respective one of the plurality of ML models trained with that subset of the training data.

11. The system of claim 7, wherein one of the plurality of ML models is trained by one of the subsets of the training data that is associated with a mix of the distinct scenarios.

12. The system of any of claims 1 to 11, wherein the decoder satisfies a specification defining a model type, a model structure, an activation function, a maximum floating-point operations per seconds (FLOPs), or a maximum number or size of parameters.

13. A UE, comprising the encoder of claim 1.

14. A gNB of a cellular network, comprising the decoder of claim 1.

15. A method, performed by test equipment, comprising: obtaining training data associated with channel state information (CSI) and a cellular network; inputting the training data to one or more two-sided machine learning (ML) models, each comprising an encoder and a decoder; compressing, by the encoder, the CSI, resulting in compressed CSI data; and decoding, by the decoder, the compressed CSI data, to obtain second CSI data.

16. The method of claim 15, further comprising enforcing at least one of: weight constraints, dropout layers, early stopping criteria, or noise injection.

17. The method of claim 15, wherein the training data comprises data associated with a clustered delay line (CDL) channel or an Urban Macrocell (Uma) channel.

18. The method of claim 15, wherein obtaining the training data comprises: obtaining portions of the training data from a plurality of data sources, and generating synthetic data404904-9517-2718, v. 3from the portions of the training data, wherein the training data comprises a combination of the portions of the training data and the synthetic data.

19. The method of claim 18, wherein generating the synthetic data comprises applying a generative model or a statistical model to the portions of the training data to generate the synthetic data.

20. The method of claim 15, wherein the one or more two-sided ML models comprise a plurality of ML models, each trained with a respective one of a plurality of subsets of the training data.

21. The method of claim 20, wherein each of the plurality of the subsets of training data are associated with distinct scenarios comprising different channel conditions, different network conditions, or different combinations of a channel condition and network condition.

22. The method of claim 21, wherein the different network conditions comprises at least one of: an indoor propagation environment, an outdoor propagation environment, a location, or a distance.

23. The method of claim 21, wherein the different channel conditions comprises at least one of: a different doppler speed, a different carrier frequency, a different channel.

24. The method of claim 20, wherein the plurality of subsets of the training data are each associated with a training set identifier (ID) that is associated with a respective one of the plurality of ML models trained with that subset of the training data.

25. The method of claim 21, wherein one of the plurality of ML models is trained by one of the subsets of the training data that is associated with a mix of the distinct scenarios.

26. The method of claim 15, wherein the decoder satisfies a specification defining a model type, a model structure, an activation function, a maximum floating-point operations per seconds (FLOPs), or a maximum number or size of parameters.

27. The method of claim 15, further comprising testing a device under test user equipment (DUT UE) with the decoder.

28. The method of claim 27, further comprising tuning a second encoder of the DUT UE based on the compressed CSI data.414904-9517-2718, v.

329. The method of claim 27, wherein testing the DUT UE with the decoder comprises selecting the decoder based on test data that is provided as input to the DUT UE.

30. A non-transitory computer readable medium storing instructions that, when executed by one or more processing devices, to cause the one or more processing devices to perform a method according to any of claims 15 to 29.4904-9517-2718, v. 3

Images